DuoNeural G-TAP v3 Quantization: LiquidAI/LFM2.5-350M (Unified Repository)
[!NOTE] Experimental Release: Pending Further Verification / Empirical Validation
Quantized using Generalized Thouless-Anderson-Palmer (G-TAP v3) non-equilibrium cavity field conditioning and TAP-DPQ dynamic programming codebook optimization developed by the DuoNeural Research Lab (Jesse Caldwell, Archon, and Aura ✨).
Model Architecture & DuoNeural Innovation
LiquidAI/LFM2.5-350M is a compact hybrid foundation model engineered for edge platforms, mobile NPUs, and local agents:
- Topology: 16 total decoder layers (10 double-gated short convolution blocks interleaved with 6 GQA blocks).
- Convolution Mechanics: Finite-Impulse-Response (FIR) causal depthwise 1D convolutions ($K=3$) with data-dependent multiplicative input gating ($1024 \to 3072$ in-projection).
- Attention Mechanism: Periodic Grouped-Query Attention (16 query heads, 8 key-value heads, head dimension 64).
- Context Capacity: 32,768 tokens native context length over a 65,536 vocabulary.
The G-TAP v3 Advantage on Hybrid Convolutional Backbones
Standard post-training quantization destroys the dynamic range of double-gated convolutions because discretization error multiplies across the gate and convolution channels ($\epsilon_{\text{gate}} \cdot \epsilon_{\text{conv}}$).
G-TAP v3 explicitly calculates and subtracts the Onsager cavity back-reaction field: $$\delta W_{ij} = -\frac{\beta}{1 + \beta \Omega_i} \Theta_{ij}$$ acting as an active thermodynamic noise filter that isolates the mutual feedback between the dynamic gating projection and the rolling FIR ring buffer. This stabilizes intermediate representations and prevents activation saturation across all 16 layers.
Head-to-Head Comparative Benchmark Matrix
Benchmarked on NVIDIA GeForce RTX 4080 Super (32GB VRAM) across 25 GSM8K multi-step math problems, 10 Python AST code problems, and 5 Hermes structured tool calling scenarios.
| Checkpoint File | Target BPW | Size | Perplexity | GSM8K Accuracy | Python Code AST | Hermes Tool AST | Decode Speed |
|---|---|---|---|---|---|---|---|
| LFM2.5-350M-G-TAP-v3-Q4_K_M.gguf | ~4.50 bpw |
218.7 MiB | N/A | 72.0% | 60.0% | 100.0% | 820.1 t/s |
| LFM2.5-350M-G-TAP-v3-IQ3_XXS.gguf | ~3.06 bpw |
151.1 MiB | N/A | 16.0% | 0.0% | 100.0% | 854.6 t/s |
| LFM2.5-350M-G-TAP-v3-IQ2_M.gguf | ~2.70 bpw |
136.3 MiB | N/A | 0.0% | 10.0% | 20.0% | 783.8 t/s |
| LFM2.5-350M-G-TAP-v3-IQ2_XXS.gguf | ~2.06 bpw |
118.0 MiB | N/A | 0.0% | 0.0% | 0.0% | 760.7 t/s |
Checkpoint Selection Guide
LFM2.5-350M-G-TAP-v3-Q4_K_M.gguf(~4.50 bpw): Maximum structural fidelity.LFM2.5-350M-G-TAP-v3-IQ3_XXS.gguf(~3.06 bpw): Sweet-spot edge quant (~150 MB).LFM2.5-350M-G-TAP-v3-IQ2_M.gguf(~2.70 bpw): Ultra-compact sub-2.7-bit quant for microcontrollers and edge NPUs.LFM2.5-350M-G-TAP-v3-IQ2_XXS.gguf(~2.06 bpw): Extreme compression for ultra-low memory footprints.
Quickstart & Usage
Running with llama.cpp
./llama-cli -m LFM2.5-350M-G-TAP-v3-Q4_K_M.gguf -p "<|im_start|>user\nWrite a python function to compute the Fibonacci sequence.<|im_end|>\n<|im_start|>assistant\n" -n 256 -ngl 99
Ollama Modelfile
FROM ./LFM2.5-350M-G-TAP-v3-Q4_K_M.gguf
TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
PARAMETER temperature 0.6
Citation & Authorship
@misc{duoneural2026gtap_lfm350m,
author = {Jesse Caldwell and Archon and Aura ✨},
title = {Generalized Thouless-Anderson-Palmer (G-TAP v3) Non-Equilibrium Cavity Quantization on Liquid AI LFM2.5 Architectures},
year = {2026},
publisher = {DuoNeural Research Lab},
howpublished = {\url{https://huggingface.co/DuoNeural/LFM2.5-350M-GTAP-v3-GGUF}}
}
DuoNeural Cognitive Light Cone — Distributed Neuro-Symbiotic System
Jesse Caldwell, Archon, Aura ✨